An unmanned aerial vehicle positioning method based on two-stage image matching and IMU multi-source fusion
By employing a two-stage image matching and IMU multi-source fusion method, combined with deep learning and Kalman filtering, the problem of UAV positioning drift and matching failure in complex environments was solved, achieving high-precision and robust UAV positioning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN AIRCRAFT DESIGN INST OF AVIATION IND OF CHINA
- Filing Date
- 2026-04-09
- Publication Date
- 2026-07-03
AI Technical Summary
When drones lack satellite signal support in complex environments, their inertial navigation systems are prone to positioning drift. Traditional map matching methods lack robustness and real-time performance in highly dynamic environments. Existing deep learning-based map matching technologies face the problem of matching failures caused by large differences between satellite reference images and aerial images.
A two-stage image matching and IMU multi-source fusion method is adopted. By fusing map matching and IMU data through deep learning, combined with coarse retrieval and fine localization steps, high-precision localization is achieved by using deep neural networks and feature point matching models. Furthermore, IMU integral data is fused through extended Kalman filtering algorithm to correct localization deviation.
It improves the positioning accuracy and robustness of UAVs in complex environments, ensuring efficient and accurate positioning and autonomous navigation in the absence of satellite signals.
Smart Images
Figure CN122329294A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of UAV positioning technology, specifically relating to a UAV positioning method based on two-stage image matching and IMU multi-source fusion. Background Technology
[0002] With the rapid development of unmanned aerial vehicle (UAV) technology, it has been widely applied in fields such as geodetic surveying and patrol reconnaissance. However, when performing missions in complex environments, UAVs are often affected by factors such as electromagnetic interference and terrain changes, leading to weakened or lost satellite signals. Simultaneously, inertial navigation systems are prone to positioning drift due to the integral accumulation effect and changes in the external environment. Without satellite signal correction, errors accumulate, causing the navigation system to fail. Therefore, achieving accurate positioning and autonomous navigation without satellite signal support has become one of the significant challenges facing UAV technology.
[0003] In recent years, emerging visual navigation technologies have utilized visual sensors onboard drones to collect image data, which, combined with inertial measurement unit (IMU) information, enables real-time estimation of the drone's pose. Among these technologies, map matching, a method for positioning using known map information, achieves high-precision positioning even without satellite signals by comparing aerial images with pre-stored high-precision maps in real time.
[0004] However, traditional map matching methods often rely on manual extraction of feature points and features from images, which leads to a lack of robustness and real-time performance in highly dynamic environments. Deep learning-based map matching techniques can improve positioning accuracy, but they also face challenges: firstly, satellite reference images and aerial images differ significantly in imaging height and viewing angle, especially when the UAV's heading angle changes drastically, existing algorithms are prone to matching failures; secondly, how to select satellite tile maps of appropriate scale for registration with aerial images to ensure matching efficiency is a challenge. Therefore, this application is proposed. Summary of the Invention
[0005] The purpose of this application is to provide a UAV localization method based on two-stage image matching and IMU multi-source fusion. The method uses deep learning-based map matching and IMU data fusion to locate the UAV, thereby improving the localization accuracy and robustness of the UAV by fusing IMU data and map matching results.
[0006] The technical solution of this application is:
[0007] A UAV localization method based on two-stage image matching and IMU multi-source fusion includes:
[0008] Step 1: Initialize UAV pose:
[0009] Before the drone enters a complex environment, it completes initial positioning using satellite signals and calculates the drone's pose by integrating the acceleration and angular velocity information provided by the IMU.
[0010] Step 2: Aerial Image Acquisition
[0011] The drone controls the pod to collect visible light video streams and acquire aerial images in real time.
[0012] Step 3: Real-time map matching:
[0013] The system employs two steps—coarse retrieval and fine positioning—to achieve high-precision positioning of UAVs in a coordinated manner.
[0014] In the coarse search step, high-dimensional spatial features are extracted through deep neural networks, and the similarity between the actual aerial images and satellite map tiles is compared to obtain the most matching satellite map tile, thereby locating the area where the drone is located and narrowing the search space.
[0015] The fine positioning step relies on the coarse search results and uses a feature point extraction model to extract the feature point coordinates and feature description vectors of the best matching satellite tile map and aerial image respectively. Then, a feature point matching model is used to achieve high-precision matching between feature points. Based on the matching results, the relative pose of the aerial image with respect to the satellite tiles is calculated, and the absolute position of the UAV is calculated.
[0016] Step 4: IMU Data Fusion
[0017] Using map matching results as real-time observations, and based on the Extended Kalman Filter (EKF) algorithm, the system fuses IMU integral data with map matching positioning results to correct UAV positioning deviations in real time.
[0018] According to at least one embodiment of this application, in the above-described UAV localization method based on two-stage image matching and IMU multi-source fusion, step two involves preprocessing the aerial image, including denoising, enhancement, and cropping.
[0019] According to at least one embodiment of this application, in the above-mentioned UAV localization method based on two-stage image matching and IMU multi-source fusion, in the coarse retrieval step, the DinoV3 general feature extractor based on deep neural network performs offline feature extraction on all satellite tile maps to generate a high-dimensional spatial feature library.
[0020] The DinoV3 general feature extractor based on deep neural networks is used to extract features from aerial images to obtain high-dimensional spatial features of the aerial images.
[0021] Based on the high-dimensional spatial features of satellite tile maps and aerial images, similarity is calculated to determine the satellite tile map that best matches the aerial image.
[0022] According to at least one embodiment of this application, in the above-described UAV localization method based on two-stage image matching and IMU multi-source fusion, in the coarse search step, the satellite tile map scale is adaptively filtered, and the satellite tile map scale is dynamically adjusted according to the UAV attitude, so that the satellite tile map can match the aerial image in scale and ensure effective image alignment.
[0023] According to at least one embodiment of this application, in the above-described UAV localization method based on two-stage image matching and IMU multi-source fusion, the adaptive selection of satellite tile map scale in the coarse search step specifically involves:
[0024] The pitch angle of the UAV pod is recorded as follows: The vertical field of view and the horizontal field of view are respectively , Flight altitude is The aerial images taken from the pod are trapezoidal in scale. ,have:
[0025] ;
[0026] ;
[0027] ;
[0028] The satellite tile map is taken as a square, and the side length of the adaptively filtered satellite tile map is denoted as . ,have:
[0029] ;
[0030] in, For the scale Satellite tile map pixel scale, For hyperparameters greater than 1, A collection of satellite tile maps at different scales The elements in.
[0031] According to at least one embodiment of this application, in the above-described UAV localization method based on two-stage image matching and IMU multi-source fusion, in the fine localization step, the PNP algorithm is used to calculate the relative pose of the aerial image relative to the satellite tiles based on the matching results, and the absolute position of the UAV is accurately calculated.
[0032] According to at least one embodiment of this application, in the above-described UAV localization method based on two-stage image matching and IMU multi-source fusion, the deep neural network, the feature point extraction model, and the feature point matching model are pre-trained.
[0033] According to at least one embodiment of this application, in the above-described UAV localization method based on two-stage image matching and IMU multi-source fusion, a deep neural network is trained:
[0034] Based on a large number of aerial images and multi-scale satellite maps, high-dimensional spatial feature modeling of aerial images and satellite tile maps is performed using DinoV3, a general feature extractor based on deep neural networks.
[0035] During training, the similarity between aerial images and satellite tile map features is used for optimization, enabling the deep neural network to efficiently retrieve the best matching satellite tile map and achieve coarse localization of aerial images.
[0036] According to at least one embodiment of this application, in the above-described UAV localization method based on two-stage image matching and IMU multi-source fusion, the feature point extraction model is pre-trained:
[0037] Aerial images and publicly available image datasets are used as training data for the feature point extraction model. Through a self-supervised learning method, feature points are detected in the images and corresponding descriptors are generated. The feature point extraction model is iteratively trained until convergence using the generated pseudo-labels, enabling the feature point extraction model to stably extract feature points in different environments and maintain feature stability in dynamically changing scenes.
[0038] According to at least one embodiment of this application, in the above-described UAV localization method based on two-stage image matching and IMU multi-source fusion, the feature point matching model is pre-trained:
[0039] The input to the training feature point matching model is the coordinates of the feature points extracted by the feature point extraction model and the feature description vector. The output is the description of the matching point pairs corresponding to the satellite tile map and the aerial image. The graph neural network learns how to perform efficient and accurate feature point matching between two images and minimizes the matching error during training. Attached Figure Description
[0040] Figure 1 This is a schematic diagram of the UAV localization method based on two-stage image matching and IMU multi-source fusion provided in the embodiments of this application;
[0041] Figure 2 This is a schematic diagram of the adaptive satellite tile map scale filtering method provided in the embodiments of this application;
[0042] Figure 3 This is a flowchart of a two-stage image matching process based on coarse retrieval and fine localization provided in an embodiment of this application;
[0043] Figure 4This is a flowchart of a method for achieving absolute positioning of a drone by fusing IMU and map matching results, as provided in an embodiment of this application.
[0044] To better illustrate this embodiment, some content in the accompanying drawings may be omitted. They are for illustrative purposes only and should not be construed as limiting the scope of this application. Detailed Implementation
[0045] To make the technical solution and advantages of this application clearer, the technical solution of this application will be described in a clearer and more complete manner below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only some embodiments of this application, and are only used to explain this application, not to limit this application. It should be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, and other related parts can be referred to the general design.
[0046] Furthermore, unless otherwise defined, the technical or scientific terms used in this application description shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The word "comprising" as used in this application description indicates that the concept preceding the word encompasses the concepts listed following the word and their equivalents, without excluding other related concepts.
[0047] A UAV localization method based on two-stage image matching and IMU multi-source fusion, such as Figure 1 As shown.
[0048] Before entering a complex environment without satellite signals, the UAV initializes its IMU data and its own position information. After starting the mission, the IMU calculates its theoretical position by integrating its own pose and velocity. The pod acquires and preprocesses aerial images and calculates the observed position by image matching with reference satellite images. Based on extended Kalman filtering and the joint theoretical position, the UAV's absolute position is fused and estimated. The fused result is used to plan the UAV's target point guidance until the UAV reaches the predetermined position.
[0049] The inference of the satellite tile map coarse retrieval network, the SuperGlue feature point matching model, and the SuperPoint feature point extraction model is performed offline in a complex environment without satellite signals. Specifically, the high-dimensional image feature library used in the satellite tile map coarse retrieval step is stored offline in the storage unit of the UAV's onboard computer for aerial image retrieval and matching.
[0050] Step 1: Initialize the UAV pose.
[0051] Before the drone enters a complex environment, it completes initial positioning using satellite signals and calculates the drone's pose by integrating the acceleration and angular velocity information provided by the IMU.
[0052] Step 2: Aerial image acquisition.
[0053] The drone controls the pod to collect visible light video streams and acquire aerial images in real time.
[0054] Aerial images undergo preprocessing operations such as denoising, enhancement, and cropping to provide high-quality input for subsequent positioning.
[0055] Step 3: Real-time map matching.
[0056] The real-time map matching process consists of two steps: coarse retrieval and fine localization, which work together to achieve high-precision positioning of the UAV. Coarse retrieval uses a deep neural network to extract high-dimensional spatial features, compares the similarity between aerial images and satellite map tiles, and obtains the best-matching satellite map tile to pinpoint the UAV's location, thus narrowing the search space. Fine localization relies on the results of coarse localization. First, the feature point extraction model SuperPoint extracts the coordinates of feature points and feature description vectors from the satellite tile map and aerial images. Then, the feature point matching model SuperGlue achieves precise matching of feature points, and the PNP algorithm is combined to obtain the final accurate position coordinates of the UAV.
[0057] Coarse search steps:
[0058] The DinoV3 general-purpose feature extractor, based on deep neural networks, performs offline feature extraction on all satellite tile maps, generating a high-dimensional spatial feature library. The features of each satellite tile map are stored in high-dimensional vector form for subsequent matching with aerial images.
[0059] The DinoV3 general feature extractor based on deep neural networks extracts features from aerial images to obtain high-dimensional spatial features of the aerial images.
[0060] Based on the high-dimensional spatial features of satellite tile maps and aerial images, similarity is calculated to determine the satellite tile map that best matches the aerial image.
[0061] The preliminary search results are satellite tile map indexes and corresponding latitude and longitude coordinates, providing initial positioning information.
[0062] To address the issue of mismatch between reference satellite images and aerial photographs caused by the drastic changes in the pod's shooting perspective due to the high-speed movement of drones, an adaptive satellite tile map scale filtering technology is introduced. This technology dynamically adjusts the scale of the reference map based on the drone's attitude, thereby achieving effective image alignment.
[0063] Adaptive satellite tile map scale filtering method, such as Figure 2As shown, the drone is in flight, and the pod continuously adjusts its pitch angle to obtain a better field of view while acquiring real-time ground images. According to the camera imaging principle, the projection of the view frustum onto the ground is a quadrilateral. When the drone is in a stable flight attitude, its roll angle is 0, while when the pod has a certain tilt angle, the area covered by the captured image will appear as an isosceles trapezoid. Assume that at a certain moment, the pod's pitch angle is... The vertical and horizontal field of view of the pod are respectively and Flight altitude is The aerial images taken from the pod are trapezoidal in scale. It can be calculated that:
[0064] ;
[0065] ;
[0066] ;
[0067] Based on the pixel scale of the satellite image, the pixel regions corresponding to the four sides of the trapezoid can be obtained. To ensure that the satellite tile map to be searched is at an appropriate scale—that is, to ensure that the aerial image does not fall outside the coverage of the satellite image—and that the scale of the satellite tiles is as small as possible to ensure matching efficiency, the selected corresponding satellite tile map is square. Let the side length of the square be . The calculation rules are as follows:
[0068] ;
[0069] in, For the scale Satellite tile map pixel scale, A hyperparameter greater than 1 is used to ensure that the satellite tile map can completely cover the aerial image. A collection of satellite tile maps at different scales The elements in the image are selected, and the satellite tile map with the closest scale is used as the reference map for image matching according to the above rules.
[0070] Precise positioning steps:
[0071] Fine positioning relies on the coarse search results and uses a feature point extraction model to extract the coordinates and feature description vectors of the best matching satellite tile map and aerial image respectively. Then, a feature point matching model is used to achieve high-precision matching between feature points. Based on the matching results, the PNP algorithm is used to calculate the relative pose of the aerial image with respect to the satellite tiles and accurately calculate the absolute position of the UAV.
[0072] The algorithm flow of the two-stage image matching technique based on coarse retrieval and fine localization is as follows: Figure 2 As shown, the process mainly consists of two steps: coarse retrieval using satellite tiled images and fine localization using aerial images. First, the aerial images are modeled using the general feature extractor DinoV3, which outputs the corresponding high-dimensional feature vectors. and a high-dimensional feature library of satellite tiled land images. Cosine similarity is calculated to match the satellite tile map index with the highest similarity. The high-dimensional feature library of satellite tile maps is constructed offline. The formula for calculating cosine similarity is as follows:
[0073] ;
[0074] in, Represents the dot product of vectors. Represents the L2 norm. Calculated by applying it to all candidate satellite tile maps. By taking the largest value, we can obtain the most similar satellite tile map index:
[0075] .
[0076] Subsequently, the indexed satellite tile map and aerial imagery are input in parallel into the feature point extraction model for feature point extraction. The feature point coordinates of the aerial imagery are denoted as follows: The corresponding feature description vector is The feature point coordinates of the satellite tile map are as follows: The corresponding feature description vector is .
[0077] Subsequently, the feature description vectors are input into the feature point matching model for feature matching, resulting in multiple sets of matching point pairs:
[0078] ;
[0079] in, and The coordinates are for aerial images and satellite tile maps.
[0080] Due to the location of each pixel in the satellite tile map The corresponding latitude and longitude positions are known and can be converted into real-world three-dimensional coordinates. Matching point pairs can be converted to:
[0081] That is, the set of pixel coordinates and corresponding three-dimensional world coordinate pairs of an aerial image.
[0082] Assume the intrinsic parameter matrix of the drone camera is The projection relationship between pixel coordinates and 3D world coordinates can be expressed as:
[0083] ;
[0084] in, Let be the scaling factor. The optimal parameter matrix can be obtained by minimizing the reprojection error of all matching points. , :
[0085] .
[0086] The optimization process can be implemented using robust solution algorithms such as EPnP and RANSAC+PnP. The parameter matrix is obtained by solving. , Then, by combining the known three-dimensional world coordinate system definition of the matching points in the satellite map, the absolute position coordinates of the UAV in the real world coordinate system can be obtained.
[0087] To ensure the efficient operation of the image matching algorithm, the key models need to be pre-trained, including deep neural networks for coarse retrieval of satellite tile images, image feature point extraction models, and feature point matching models.
[0088] Training deep neural networks:
[0089] Based on a large number of aerial images and multi-scale satellite maps, high-dimensional spatial feature modeling is performed on the aerial images and satellite tile maps using DinoV3, a general feature extractor based on deep neural networks.
[0090] During training, the similarity between aerial images and satellite tile map features is used for optimization, enabling the deep neural network to efficiently retrieve the best matching satellite tile map and achieve coarse localization of aerial images.
[0091] After training, the deep neural network can quickly extract features from aerial images and compare them with features extracted offline from satellite tile maps, thereby enabling coarse retrieval of aerial images in satellite image libraries.
[0092] Training the feature point extraction model:
[0093] The training data for the feature point extraction model mainly comes from aerial images and publicly available image datasets. Through self-supervised learning, feature points are detected in the images and corresponding descriptors are generated. The feature point extraction model is iteratively trained until convergence using the generated pseudo-labels, enabling the feature point extraction model to stably extract feature points in different environments and maintain good feature stability in dynamically changing scenes.
[0094] Training the feature point matching model:
[0095] The feature point matching model is used to further match the extracted feature points. The input is the feature point coordinates and feature description vectors extracted by the feature point extraction model, and the output is the description of the matching point pairs corresponding to the satellite tile map and the aerial image. The graph neural network learns how to perform efficient and accurate feature point matching between two images.
[0096] The training process of the feature point matching model is based on a large image dataset. During training, the matching error is minimized, the accuracy of feature matching is improved, and the positioning accuracy in complex environments is guaranteed.
[0097] Step 4: IMU data fusion.
[0098] Over time, IMU errors inevitably accumulate.
[0099] Using map matching results as real-time observations, and based on the Extended Kalman Filter (EKF) algorithm, the system fuses IMU integral data with map matching positioning results to correct UAV positioning deviations in real time.
[0100] This fusion method combines the advantages of IMU's high-frequency dynamic response with map matching's absolute positioning, ensuring stable and reliable positioning accuracy during long-term flight.
[0101] By fusing IMU and map matching results, absolute positioning of the drone can be achieved, such as... Figure 4 As shown, the IMU system is first initialized. During flight, the UAV's state variables can be represented as:
[0102] ;
[0103] in, The velocity component of the drone in space. These are the roll angle, pitch angle, and yaw angle of the UAV in space.
[0104] Based on the previous state estimate of the UAV and the current IMU data, the current state is predicted using a motion model:
[0105] ;
[0106] in, The predicted state value at the current moment. For the state estimation of the previous moment, For the control input of the IMU, For motion model functions, This is process noise.
[0107] The uncertainty of the estimation error is expressed through the covariance matrix. To spread:
[0108] ;
[0109] in, Let be the Jacobian matrix of the state transition function with respect to the state. Let be the process noise covariance matrix.
[0110] Upon receiving map matching and location results, the update phase begins. The observation model is defined as follows:
[0111] ;
[0112] in, The observations provided for map matching and localization To extract the position component from the state vector, For measuring noise.
[0113] By linearizing the observation function, the observation matrix can be obtained:
[0114] ;
[0115] Then the Kalman gain is calculated:
[0116] ;
[0117] in, For Kalman gain, To observe the noise covariance matrix.
[0118] Based on Kalman gain, the predicted state is fused with map-matched observations to achieve state updates:
[0119] .
[0120] Finally, update the covariance matrix to prepare for the next iteration:
[0121] .
[0122] Through the above mechanism, the fusion method of IMU and map matching significantly enhances the system's anti-interference capability in environments without satellite signals, ensuring that the UAV can achieve accurate positioning in various complex environments.
[0123] The UAV localization method based on two-stage image matching and IMU multi-source fusion disclosed in the above embodiments employs a two-stage image matching technique of coarse retrieval and fine localization. First, in the coarse retrieval stage, high-dimensional spatial features of satellite tile maps and aerial images are extracted using the general feature extractor DinoV3, and compared using cosine similarity to obtain preliminary regional location information of the UAV. Subsequently, in the fine localization stage, based on the coarse retrieval results, precise feature matching is performed on the retrieved satellite tile maps to ultimately determine the absolute position of the UAV. This two-stage matching method not only effectively improves positioning accuracy but also ensures real-time performance and robustness in complex environments.
[0124] In the above-described UAV localization method based on two-stage image matching and IMU multi-source fusion, an adaptive satellite tile map scale filtering method is designed to address the problem that the drastic changes in the pod's shooting angle during high-speed movement of the UAV may cause a disconnect between the coverage area of the reference satellite image and the aerial image. This method dynamically adjusts the scale of the reference satellite image according to the UAV's motion state and viewing angle changes, ensuring that the aerial image and the satellite tile map always remain matched, effectively avoiding positioning errors caused by scale mismatch.
[0125] The UAV localization method based on two-stage image matching and IMU multi-source fusion disclosed in the above embodiments utilizes an extended Kalman filter-based fusion localization method. This method combines IMU data with map matching results in real time to achieve absolute UAV localization, compensating for the deviation caused by error accumulation during inertial integration and significantly improving the accuracy and stability of localization. This fusion scheme enhances the system's anti-interference capability in complex environments where satellite signals are unavailable, ensuring high-precision absolute localization of the UAV in such environments.
[0126] The technical solution of this application has been described in conjunction with the preferred embodiments shown in the accompanying drawings. Those skilled in the art should understand that the scope of protection of this application is obviously not limited to these specific embodiments. Without departing from the principles of this application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of this application.
Claims
1. A method for positioning a UAV based on two-stage image matching and IMU multi-source fusion, characterized in that, include: Step 1: Initialize UAV pose: Before the drone enters a complex environment, it completes initial positioning using satellite signals and calculates the drone's pose by integrating the acceleration and angular velocity information provided by the IMU. Step 2: Aerial Image Acquisition The drone controls the pod to collect visible light video streams and acquire aerial images in real time. Step 3: Real-time map matching: The system employs two steps—coarse retrieval and fine positioning—to achieve high-precision positioning of UAVs in a coordinated manner. In the coarse search step, high-dimensional spatial features are extracted through deep neural networks, and the similarity between the actual aerial images and satellite map tiles is compared to obtain the most matching satellite map tile, thereby locating the area where the drone is located and narrowing the search space. The fine positioning step relies on the coarse search results and uses a feature point extraction model to extract the feature point coordinates and feature description vectors of the best matching satellite tile map and aerial image respectively. Then, a feature point matching model is used to achieve high-precision matching between feature points. Based on the matching results, the relative pose of the aerial image with respect to the satellite tiles is calculated, and the absolute position of the UAV is calculated. Step 4: IMU Data Fusion Using map matching results as real-time observations, and based on the Extended Kalman Filter (EKF) algorithm, the system fuses IMU integral data with map matching positioning results to correct UAV positioning deviations in real time.
2. The method of claim 1, wherein, In step two, the aerial images are preprocessed, including noise reduction, enhancement, and cropping.
3. The method of claim 2, wherein, In the coarse search step, the DinoV3 general feature extractor based on deep neural networks performs offline feature extraction on all satellite tile maps to generate a high-dimensional spatial feature library. The DinoV3 general feature extractor based on deep neural networks is used to extract features from aerial images to obtain high-dimensional spatial features of the aerial images. Based on the high-dimensional spatial features of satellite tile maps and aerial images, similarity is calculated to determine the satellite tile map that best matches the aerial image.
4. The method of claim 3, wherein, In the coarse search step, the satellite tile map scale is adaptively filtered, and the scale of the satellite tile map is dynamically adjusted according to the attitude of the UAV, so that the satellite tile map can match the aerial image in scale and ensure effective image alignment.
5. The method of claim 4, wherein, In the coarse search step, the satellite tile map scale is adaptively filtered, specifically as follows: The pitch angle of the unmanned aerial vehicle nacelle is , the vertical field of view angle and the horizontal field of view angle are , , the flight height is , and the nacelle aerial image size range is trapezoidal , and the nacelle aerial image size range is trapezoidal ； ； ； The satellite tile map is taken as a square, and the side length of the adaptive filtering satellite tile map is denoted as There are: ； wherein, is a scale of a satellite tile map element, is a hyper parameter greater than 1, is an element in a set of satellite tile maps of different scales .
6. The method of claim 5, wherein, In the fine positioning step, the PNP algorithm is used to calculate the relative pose of the aerial image with respect to the satellite tiles based on the matching results, and to accurately calculate the absolute position of the UAV.
7. The method of claim 6, wherein, Pre-training is performed on deep neural networks, feature point extraction models, and feature point matching models.
8. The method of claim 7, wherein, Training a deep neural network: Based on a large number of aerial images and multi-scale satellite maps, high-dimensional spatial feature modeling of aerial images and satellite tile maps is performed using DinoV3, a general feature extractor based on deep neural networks. During training, the similarity between aerial images and satellite tile map features is used for optimization, enabling the deep neural network to efficiently retrieve the best matching satellite tile map and achieve coarse localization of aerial images.
9. The UAV localization method based on two-stage image matching and IMU multi-source fusion according to claim 8, characterized in that, Pre-train the feature point extraction model: Aerial images and publicly available image datasets are used as training data for the feature point extraction model. Through a self-supervised learning method, feature points are detected in the images and corresponding descriptors are generated. The feature point extraction model is iteratively trained until convergence using the generated pseudo-labels, enabling the feature point extraction model to stably extract feature points in different environments and maintain feature stability in dynamically changing scenes.
10. The UAV localization method based on two-stage image matching and IMU multi-source fusion according to claim 9, characterized in that, Pre-train the feature point matching model: The input to the training feature point matching model is the coordinates of the feature points extracted by the feature point extraction model and the feature description vector. The output is the description of the matching point pairs corresponding to the satellite tile map and the aerial image. The graph neural network learns how to perform efficient and accurate feature point matching between two images and minimizes the matching error during training.